[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M79B0EPDDNQC6WGJ87FBYZ2S":3,"skill-same:01M79B0EPDDNQC6WGJ87FBYZ2S":53},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":27,"repositoryHtmlUrl":28,"repositoryHomepageUrl":6,"ownerHtmlUrl":29,"ownerBlogUrl":30,"ownerDescription":31,"updatedAt":32,"install":33,"resources":46},"01M79B0EPDDNQC6WGJ87FBYZ2S","review",null,"Reviews a capability as architecture and then as prose — traces its flows and decision points, maps its units against the resources powering them, diagrams it, stress-tests the logic, and checks that every command it documents exists and every artifact it writes has a reader. Use when asked to review, audit, critique or sanity-check a capability or its skills, before contributing one upstream, after building or evolving one, when something in a capability \"looks wrong\" but nobody can say where, or when a capability's prose and its tool may have drifted apart. Read-only, so it reports findings and fixes nothing: Do NOT use to apply a fix — that is capability-contribute upstream, or capability-evolve for one user.","AlmogBaku\u002Faos","AlmogBaku","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F98982?v=4","01M79AYYMZE5E1BEVEP6P829C4","Almog Baku","01M79AYYP0D19RBK33EBYYR3F8",5,"MIT","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Fcapabilities\u002Fcapability-lifecycle\u002Fskills\u002Freview\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: review\ndescription: \"Reviews a capability as architecture and then as prose — traces its flows and decision points, maps its units against the resources powering them, diagrams it, stress-tests the logic, and checks that every command it documents exists and every artifact it writes has a reader. Use when asked to review, audit, critique or sanity-check a capability or its skills, before contributing one upstream, after building or evolving one, when something in a capability \\\"looks wrong\\\" but nobody can say where, or when a capability's prose and its tool may have drifted apart. Read-only, so it reports findings and fixes nothing: Do NOT use to apply a fix — that is capability-contribute upstream, or capability-evolve for one user.\"\n---\n\n# capability-review\n\nNot in context yet? Load the `capability-lifecycle` skill first — the map, the contract, and\nthe naming rules. **Read-only** counterpart to {{skill: build}}: writes nothing, reports\neverything for a human to act on.\n\n**A capability is software.** Its components are skills, agents, tools, crons and overlays; its\ninterfaces are prose an LLM follows; its bugs are unreachable branches, dangling edges and\nunstated preconditions.\n\nReviewing it sentence-by-sentence finds typos. Reviewing it as a *system* finds the defects that\nmatter — **which is why steps 1–5 come first and the textual sweep comes last, not the other way\nround.** Two real runs of this skill on this kit each found every high-severity defect in steps\n1–5 and none in the sweep, and one of those was a data-loss bug that two careful readings had\nalready missed. If the cheap mechanical checks feel more productive to start with, that is the\ntemptation this order exists to defeat.\n\nSection headings carry **[A]** where the work is judgment and **[D]** where a command answers.\nThe [D] parts cannot be faked; do those literally rather than concluding them.\n\nCopy this checklist and work it in order:\n\n```\n- [ ] 1. Flows      — the journey, and every decision point\n- [ ] 2. Components — units vs the resources powering them\n- [ ] 3. Diagram    — if there is more than one moving part\n- [ ] 4. Stress     — walk the tree adversarially\n- [ ] 5. Journey    — happy path, failure modes, how each surfaces\n- [ ] 6. Sweep      — commands, readers, claims, weaker models\n- [ ] 7. Report     — findings, severity-ordered, with evidence\n```\n\n## 1. Flows and decision points **[A]**\n\nTrace the journey end to end: where does input enter, what transforms it, where does it come\nto rest. Then name the decision points — every place the SOP branches — **including the ones\nthe prose leaves implicit**. An unstated branch is where a weaker model improvises, and\nimprovisation in a capability is indistinguishable from a bug.\n\nRead `CAPABILITY.md` for what is declared, then each skill for what actually happens. Where\nthey disagree, the skill wins at runtime and the manifest is the defect.\n\n**A shipped tool's source is in scope**, not just its `--help`. Prose can only promise; the code\nis what happens. Reviews of this kit have found their worst defects — data loss, a silently\nflipped record — by reading the tool, never by reading about it.\n\n## 2. Components: units vs resources **[A]**\n\nTwo layers, kept separate — conflating them hides both kinds of gap:\n\n- **Business-logic units** — the jobs this capability does.\n- **Resources** — what powers each job: a skill, an agent, a tool, a cron, a template, an\n  overlay.\n\nThen the edges: who invokes whom, who writes what, who owns which path. **A unit with no\nresource is a wish; a resource with no unit is dead weight.** Both are findings.\n\nCheck the composition boundary too: a capability reaches another through the shared `main`\nagent, a tool on PATH, or a declared `uses` grant — a foreign skill a non-main agent loads\nwith no `uses` entry is a defect, not a style choice. Check the other direction as well:\nevery `uses` agent actually loads that skill, and no briefing places anything the manifest\ndoes not declare (a profile, an agent, a link) — the installer follows declarations only.\n\n## 3. Diagram it **[A]**\n\nDraw it when there is **more than one moving part** — several skills, or an agent, or a\ncron job, or a tool. Skip it for a single-skill capability with nothing else: a two-box diagram\ncarries no information, and shipping one teaches the reader that these diagrams are decoration —\nthe write-only artifact 6b exists to catch. The judgment is whether *relationships* matter here,\nnot how many files there are.\n\nMermaid, not ASCII — it is the kit's convention, renders on GitHub, and survives editing\nwhere hand-aligned ASCII rots on the first change. Two shapes carry most capabilities: a\n`sequenceDiagram` for the journey (who calls whom, in order) and a `flowchart` for the\ncomponent graph, resources in subgraphs.\n\n**Derive the graph from the skills and the tool first, then diff it against any diagram the\ncapability already ships — a mismatch is a finding, and the shipped one is usually the stale\nside.** Reading an existing diagram first tells you what someone believed, which is exactly what\nyou are testing. For shape only (not content), `\u003Caos_home>\u002Fupstream\u002Fcapabilities\u002Fkb\u002Fdocs\u002Fdesign.md`\nis a worked example.\n\nDrawing is the point, not decoration: a flow you cannot draw is a flow you do not understand yet,\nand ambiguity that survives prose rarely survives a diagram. The data-loss bug mentioned above\nwas found exactly here — two instructions that read as unrelated sentences named one directory,\nwhich is invisible in prose and unmissable as two nodes.\n\nTwo rules, both learned the expensive way:\n\n- A `;` inside a **sequence-diagram** message is a statement separator and breaks rendering —\n  use an em dash. It is harmless inside a quoted flowchart label.\n- **Verify by rendering, not by reading.** A block that parses in your head can still fail on\n  GitHub, and a diagram that silently fails is worse than none: the reader sees raw source and\n  stops trusting the document. Three rungs, best first: `npx @mermaid-js\u002Fmermaid-cli -i \u003Cfile>`\n  renders for real, but needs a Puppeteer browser and fails in most sandboxes and CI containers;\n  failing that, `mermaid@11`'s `parse()` under a jsdom global validates the grammar (including\n  the `;` case) with no browser; failing both, say **unverified** rather than implying it\n  renders. Name which rung you used — \"grammar-verified, not render-verified\" is an honest\n  result, \"renders correctly\" on an unrendered block is not.\n\n## 4. Stress the logic **[A]**\n\nWalk the decision tree adversarially and look for:\n\n- branches that cannot be reached, and branches that overlap so two paths both fire;\n- two paths writing the same file, or the same state, with no ordering between them;\n- an ordering assumption nobody declared (\"after the nightly steward pass\") — those rot silently\n  when the other side changes;\n- what happens on retry, on partial completion, and on a step that returns nothing.\n\nThen check soundness against whichever standing invariants this capability actually touches —\nskip the rest rather than padding: capture latency is sacred, grants default to deny, a cron job\nhas exactly one owner, one lifetime rule (`expires:`) and no second, and **attribution is the\nenforcement substrate** — every write is its own commit with the human as author and the acting\nagent as committer, so anything that lets a write reach git unattributed, or under the wrong\nsubject, breaks the audit that everything else rests on.\n\n## 5. Happy path, failure modes, journey **[A]**\n\nFor the main path and each failure mode, answer three things: what does success look like,\nhow does it fail, and **how would the failure surface** — as a crash a user reports, or as an\nexit-0 wrong answer nobody notices. The second kind is the dangerous one, and it is what\nsections 6a and 6c hunt.\n\nWhere — and only where — the capability's *triggering* is genuinely contested, propose eval\nqueries in your report; **you write no files, so do not create them.** `tests\u002Fevals\u002Fkb\u002F` shows\nthe near-miss shape, and its \"harness could not measure\" note is required reading before anyone\ntrusts a number from that harness. If triggering is not contested, say so and move on.\n\n## 6. The sweep **[D]** where a command can answer, **[A]** where judgment is needed\n\nMechanical, and it catches what careful reading misses. **This is the step to parallelise** — 6a\nand 6c are per-claim and independent, and 6c's reproductions are the slowest thing in the review,\nso hand them to subagents and collect the verdicts.\n\n**Do not parallelise steps 1–5, and do not split a capability by skill.** Every defect worth\nfinding in this kit lived *between* skills, not inside one: of ten real findings in one review,\nnine needed two or more skills in view at once and exactly one was visible in a single file. An\nagent holding only `upgrade` cannot see that its render destination is another skill's source\ndirectory — that finding *is* the relationship. The economics point the same way: the shared\ncontext every per-skill agent must re-read (manifest, entry skill, contract, naming) outweighs\nthe per-skill bodies it would save, so splitting by skill multiplies context instead of saving it.\nFan out **by capability** when reviewing several; keep one capability in one head.\n\n**a. Every command the prose names must exist, with those flags.** Run it, or read its\n`--help`. The kit's linter validates schema — frontmatter, `used_by`, reference depth — and\ncannot validate behaviour, so a verb that is really an option, a required flag omitted, or a\nflag deleted from the tool all ship green. The failure lands on the agent that followed the\ninstruction.\n\nThe kit ships the `aos_lint.gates.kb_commands` gate, but it is **hardcoded to kb and work-tracker** — run\nit only when one of those is your target. On anything else it exits 0 without opening the\ncapability under review, which reads as a pass and is not one. Otherwise extract the invocations\nyourself and diff them against `--help`. And note what no such script can see: a command can\nexist with exactly the documented flags and still be unable to do what the prose claims.\n\n**Every skill and agent the prose names must exist too.** The linter takes the mechanical half —\nslots and hardcoded names (`skills\u002Fref-dangling`, `agents\u002Fref-dangling`,\n`skills\u002Fref-hardcoded`) — so what is left for you is the *informal* reference: a rendered skill\nthat says \"hand it to the archiver\", or \"the capture skill files it\", where no such installed\nname exists. Resolve each against `aos-cap skills` and `aos-cap agents`. A reference naming\nnothing sends the agent hunting and fails silently — the model improvises something plausible\ninstead of stopping.\n\n**b. Every artifact created must name its reader.** The rule the kit already states: *a queue\nfile is only justified when the work item has no artifact of its own*. Generalised — if\nnothing reads what you write, delete the writer, not the reader. This is the single most\nrepeated defect in the kit's build-gap ledger (`\u003Caos_home>\u002Fupstream\u002Fdocs\u002FBUILD-GAPS.md`).\n\n**c. Verify guardrail claims by reproduction, never by reading.** When prose says the tool\nenforces, refuses, checks or scopes something, prove it on a throwaway fixture. Claims that a\ntool protects you are worse than absent when false, because they stop the agent from\nprotecting itself.\n\n**d. Would a weaker model reach the goal? [D] where you can prove it.** Don't conclude this —\ntest it. Take the two longest branches you drew in step 1 and name, concretely, the one fact a\nfresh agent would lack at each: an unstated prerequisite, a missing stop condition, a required\nflag the prose omits. If you cannot name one, say the branch is clean. Then check the aos-side\nmechanics: every skill declared in `used_by`, installed names computed not authored, and the\nmanifest's own counts matching the README and its prose — a version-bumping commit is exactly\nwhere those drift.\n\nFor skill quality itself, work [reference\u002Fskill-rubric.md](reference\u002Fskill-rubric.md): eight\njudgment dimensions — four for the body, four for the description — with the mechanical half\nexcluded because the linter already fails the build over it. **Read its opening rule before\nscoring anything.** The rubric is a floor: a weak dimension means something is wrong, a strong\none means stop. Polishing a skill that already reads well buys the last points by deleting\nreasoning, and reasoning is what lets a model handle the case a rule does not literally name.\nTask achievement is the target, not a tidy score.\n\n**e. Prefer a script once a defect class appears twice.** Two careful readings that both miss\nthe same kind of thing is evidence a script is cheaper than a third reading.\n\n## 7. Report **[A]**\n\nFindings severity-ordered, each with the file, what breaks, and the evidence — a command you\nran, a reproduction, a dangling edge in the graph. Include the diagram if you drew one.\n\nTwo honesty rules. **Say when a pass found nothing**, rather than manufacturing findings to\njustify the review — a clean architecture pass on a capability that has never been diagrammed\nis a real result, but so is discovering the pass was too weak to find anything. And when a\nfinding is a judgment call rather than a defect, label it as one.\n\n## Authority\n\n- Freely: read every file, run read-only commands, build throwaway fixtures in `\u002Ftmp`, run\n  `--help`, render diagrams.\n- Report-only: every finding. This skill fixes nothing.\n- Ask first: anything that would write into the capability under review, and any reproduction\n  that touches real user data — fixtures go in `\u002Ftmp`, never a live base or profile.\n","# capability-review\n\nNot in context yet? Load the `capability-lifecycle` skill first — the map, the contract, and\nthe naming rules. **Read-only** counterpart to {{skill: build}}: writes nothing, reports\neverything for a human to act on.\n\n**A capability is software.** Its components are skills, agents, tools, crons and overlays; its\ninterfaces are prose an LLM follows; its bugs are unreachable branches, dangling edges and\nunstated preconditions.\n\nReviewing it sentence-by-sentence finds typos. Reviewing it as a *system* finds the defects that\nmatter — **which is why steps 1–5 come first and the textual sweep comes last, not the other way\nround.** Two real runs of this skill on this kit each found every high-severity defect in steps\n1–5 and none in the sweep, and one of those was a data-loss bug that two careful readings had\nalready missed. If the cheap mechanical checks feel more productive to start with, that is the\ntemptation this order exists to defeat.\n\nSection headings carry **[A]** where the work is judgment and **[D]** where a command answers.\nThe [D] parts cannot be faked; do those literally rather than concluding them.\n\nCopy this checklist and work it in order:\n\n```\n- [ ] 1. Flows      — the journey, and every decision point\n- [ ] 2. Components — units vs the resources powering them\n- [ ] 3. Diagram    — if there is more than one moving part\n- [ ] 4. Stress     — walk the tree adversarially\n- [ ] 5. Journey    — happy path, failure modes, how each surfaces\n- [ ] 6. Sweep      — commands, readers, claims, weaker models\n- [ ] 7. Report     — findings, severity-ordered, with evidence\n```\n\n## 1. Flows and decision points **[A]**\n\nTrace the journey end to end: where does input enter, what transforms it, where does it come\nto rest. Then name the decision points — every place the SOP branches — **including the ones\nthe prose leaves implicit**. An unstated branch is where a weaker model improvises, and\nimprovisation in a capability is indistinguishable from a bug.\n\nRead `CAPABILITY.md` for what is declared, then each skill for what actually happens. Where\nthey disagree, the skill wins at runtime and the manifest is the defect.\n\n**A shipped tool's source is in scope**, not just its `--help`. Prose can only promise; the code\nis what happens. Reviews of this kit have found their worst defects — data loss, a silently\nflipped record — by reading the tool, never by reading about it.\n\n## 2. Components: units vs resources **[A]**\n\nTwo layers, kept separate — conflating them hides both kinds of gap:\n\n- **Business-logic units** — the jobs this capability does.\n- **Resources** — what powers each job: a skill, an agent, a tool, a cron, a template, an\n  overlay.\n\nThen the edges: who invokes whom, who writes what, who owns which path. **A unit with no\nresource is a wish; a resource with no unit is dead weight.** Both are findings.\n\nCheck the composition boundary too: a capability reaches another through the shared `main`\nagent, a tool on PATH, or a declared `uses` grant — a foreign skill a non-main agent loads\nwith no `uses` entry is a defect, not a style choice. Check the other direction as well:\nevery `uses` agent actually loads that skill, and no briefing places anything the manifest\ndoes not declare (a profile, an agent, a link) — the installer follows declarations only.\n\n## 3. Diagram it **[A]**\n\nDraw it when there is **more than one moving part** — several skills, or an agent, or a\ncron job, or a tool. Skip it for a single-skill capability with nothing else: a two-box diagram\ncarries no information, and shipping one teaches the reader that these diagrams are decoration —\nthe write-only artifact 6b exists to catch. The judgment is whether *relationships* matter here,\nnot how many files there are.\n\nMermaid, not ASCII — it is the kit's convention, renders on GitHub, and survives editing\nwhere hand-aligned ASCII rots on the first change. Two shapes carry most capabilities: a\n`sequenceDiagram` for the journey (who calls whom, in order) and a `flowchart` for the\ncomponent graph, resources in subgraphs.\n\n**Derive the graph from the skills and the tool first, then diff it against any diagram the\ncapability already ships — a mismatch is a finding, and the shipped one is usually the stale\nside.** Reading an existing diagram first tells you what someone believed, which is exactly what\nyou are testing. For shape only (not content), `\u003Caos_home>\u002Fupstream\u002Fcapabilities\u002Fkb\u002Fdocs\u002Fdesign.md`\nis a worked example.\n\nDrawing is the point, not decoration: a flow you cannot draw is a flow you do not understand yet,\nand ambiguity that survives prose rarely survives a diagram. The data-loss bug mentioned above\nwas found exactly here — two instructions that read as unrelated sentences named one directory,\nwhich is invisible in prose and unmissable as two nodes.\n\nTwo rules, both learned the expensive way:\n\n- A `;` inside a **sequence-diagram** message is a statement separator and breaks rendering —\n  use an em dash. It is harmless inside a quoted flowchart label.\n- **Verify by rendering, not by reading.** A block that parses in your head can still fail on\n  GitHub, and a diagram that silently fails is worse than none: the reader sees raw source and\n  stops trusting the document. Three rungs, best first: `npx @mermaid-js\u002Fmermaid-cli -i \u003Cfile>`\n  renders for real, but needs a Puppeteer browser and fails in most sandboxes and CI containers;\n  failing that, `mermaid@11`'s `parse()` under a jsdom global validates the grammar (including\n  the `;` case) with no browser; failing both, say **unverified** rather than implying it\n  renders. Name which rung you used — \"grammar-verified, not render-verified\" is an honest\n  result, \"renders correctly\" on an unrendered block is not.\n\n## 4. Stress the logic **[A]**\n\nWalk the decision tree adversarially and look for:\n\n- branches that cannot be reached, and branches that overlap so two paths both fire;\n- two paths writing the same file, or the same state, with no ordering between them;\n- an ordering assumption nobody declared (\"after the nightly steward pass\") — those rot silently\n  when the other side changes;\n- what happens on retry, on partial completion, and on a step that returns nothing.\n\nThen check soundness against whichever standing invariants this capability actually touches —\nskip the rest rather than padding: capture latency is sacred, grants default to deny, a cron job\nhas exactly one owner, one lifetime rule (`expires:`) and no second, and **attribution is the\nenforcement substrate** — every write is its own commit with the human as author and the acting\nagent as committer, so anything that lets a write reach git unattributed, or under the wrong\nsubject, breaks the audit that everything else rests on.\n\n## 5. Happy path, failure modes, journey **[A]**\n\nFor the main path and each failure mode, answer three things: what does success look like,\nhow does it fail, and **how would the failure surface** — as a crash a user reports, or as an\nexit-0 wrong answer nobody notices. The second kind is the dangerous one, and it is what\nsections 6a and 6c hunt.\n\nWhere — and only where — the capability's *triggering* is genuinely contested, propose eval\nqueries in your report; **you write no files, so do not create them.** `tests\u002Fevals\u002Fkb\u002F` shows\nthe near-miss shape, and its \"harness could not measure\" note is required reading before anyone\ntrusts a number from that harness. If triggering is not contested, say so and move on.\n\n## 6. The sweep **[D]** where a command can answer, **[A]** where judgment is needed\n\nMechanical, and it catches what careful reading misses. **This is the step to parallelise** — 6a\nand 6c are per-claim and independent, and 6c's reproductions are the slowest thing in the review,\nso hand them to subagents and collect the verdicts.\n\n**Do not parallelise steps 1–5, and do not split a capability by skill.** Every defect worth\nfinding in this kit lived *between* skills, not inside one: of ten real findings in one review,\nnine needed two or more skills in view at once and exactly one was visible in a single file. An\nagent holding only `upgrade` cannot see that its render destination is another skill's source\ndirectory — that finding *is* the relationship. The economics point the same way: the shared\ncontext every per-skill agent must re-read (manifest, entry skill, contract, naming) outweighs\nthe per-skill bodies it would save, so splitting by skill multiplies context instead of saving it.\nFan out **by capability** when reviewing several; keep one capability in one head.\n\n**a. Every command the prose names must exist, with those flags.** Run it, or read its\n`--help`. The kit's linter validates schema — frontmatter, `used_by`, reference depth — and\ncannot validate behaviour, so a verb that is really an option, a required flag omitted, or a\nflag deleted from the tool all ship green. The failure lands on the agent that followed the\ninstruction.\n\nThe kit ships the `aos_lint.gates.kb_commands` gate, but it is **hardcoded to kb and work-tracker** — run\nit only when one of those is your target. On anything else it exits 0 without opening the\ncapability under review, which reads as a pass and is not one. Otherwise extract the invocations\nyourself and diff them against `--help`. And note what no such script can see: a command can\nexist with exactly the documented flags and still be unable to do what the prose claims.\n\n**Every skill and agent the prose names must exist too.** The linter takes the mechanical half —\nslots and hardcoded names (`skills\u002Fref-dangling`, `agents\u002Fref-dangling`,\n`skills\u002Fref-hardcoded`) — so what is left for you is the *informal* reference: a rendered skill\nthat says \"hand it to the archiver\", or \"the capture skill files it\", where no such installed\nname exists. Resolve each against `aos-cap skills` and `aos-cap agents`. A reference naming\nnothing sends the agent hunting and fails silently — the model improvises something plausible\ninstead of stopping.\n\n**b. Every artifact created must name its reader.** The rule the kit already states: *a queue\nfile is only justified when the work item has no artifact of its own*. Generalised — if\nnothing reads what you write, delete the writer, not the reader. This is the single most\nrepeated defect in the kit's build-gap ledger (`\u003Caos_home>\u002Fupstream\u002Fdocs\u002FBUILD-GAPS.md`).\n\n**c. Verify guardrail claims by reproduction, never by reading.** When prose says the tool\nenforces, refuses, checks or scopes something, prove it on a throwaway fixture. Claims that a\ntool protects you are worse than absent when false, because they stop the agent from\nprotecting itself.\n\n**d. Would a weaker model reach the goal? [D] where you can prove it.** Don't conclude this —\ntest it. Take the two longest branches you drew in step 1 and name, concretely, the one fact a\nfresh agent would lack at each: an unstated prerequisite, a missing stop condition, a required\nflag the prose omits. If you cannot name one, say the branch is clean. Then check the aos-side\nmechanics: every skill declared in `used_by`, installed names computed not authored, and the\nmanifest's own counts matching the README and its prose — a version-bumping commit is exactly\nwhere those drift.\n\nFor skill quality itself, work [reference\u002Fskill-rubric.md](reference\u002Fskill-rubric.md): eight\njudgment dimensions — four for the body, four for the description — with the mechanical half\nexcluded because the linter already fails the build over it. **Read its opening rule before\nscoring anything.** The rubric is a floor: a weak dimension means something is wrong, a strong\none means stop. Polishing a skill that already reads well buys the last points by deleting\nreasoning, and reasoning is what lets a model handle the case a rule does not literally name.\nTask achievement is the target, not a tidy score.\n\n**e. Prefer a script once a defect class appears twice.** Two careful readings that both miss\nthe same kind of thing is evidence a script is cheaper than a third reading.\n\n## 7. Report **[A]**\n\nFindings severity-ordered, each with the file, what breaks, and the evidence — a command you\nran, a reproduction, a dangling edge in the graph. Include the diagram if you drew one.\n\nTwo honesty rules. **Say when a pass found nothing**, rather than manufacturing findings to\njustify the review — a clean architecture pass on a capability that has never been diagrammed\nis a real result, but so is discovering the pass was too weak to find anything. And when a\nfinding is a judgment call rather than a defect, label it as one.\n\n## Authority\n\n- Freely: read every file, run read-only commands, build throwaway fixtures in `\u002Ftmp`, run\n  `--help`, render diagrams.\n- Report-only: every finding. This skill fixes nothing.\n- Ask first: anything that would write into the capability under review, and any reproduction\n  that touches real user data — fixtures go in `\u002Ftmp`, never a live base or profile.\n",{"name":5,"description":7},"main","4b00482b74a48ca3adf50f02cc454094b5838334","capabilities\u002Fcapability-lifecycle\u002Fskills\u002Freview",2,"https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos","https:\u002F\u002Fgithub.com\u002FAlmogBaku","http:\u002F\u002Fwww.AlmogBaku.com","LLM hacker","2026-11-15 03:54:12",{"source":8,"skillName":5,"commands":34,"prompt":44,"zipAvailable":45},[35,38,41],{"tool":36,"command":37},"npx","npx skills add AlmogBaku\u002Faos --skill review",{"tool":39,"command":40},"bunx","bunx skills add AlmogBaku\u002Faos --skill review",{"tool":42,"command":43},"pnpm","pnpm dlx skills add AlmogBaku\u002Faos --skill review","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Fcapabilities\u002Fcapability-lifecycle\u002Fskills\u002Freview\u002FSKILL.md",true,[47],{"path":48,"resourceType":49,"resourceTypeDesc":50,"sizeBytes":51,"sourceHtmlUrl":52},"reference\u002Fskill-rubric.md","REFERENCE","参考文档",8128,"https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Fcapabilities\u002Fcapability-lifecycle\u002Fskills\u002Freview\u002Freference\u002Fskill-rubric.md",[54,61,68,75,82,89],{"skillNo":55,"name":56,"nameZh":6,"description":57,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":58,"categories":59,"keywords":60,"publishedAt":19},"01M7DZT1MJJHC87HQVK2D6RRFW","wt-update","Updates an existing commitment when the user reports progress: completion, a new deadline, a stall, or abandonment. Use when the user says \"I did that\", \"finished X\", \"still waiting on Y\", \"push that to next week\", \"I started it\", or \"forget it, I'm not doing X\". Finds the matching open action, applies the change, and routes any knowledge the report carries to the knowledge base. Do NOT use to file a new commitment (that is wt-capture) or to find time for one (wt-schedule).","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fwork-tracker\u002Fskills\u002Fwt-update\u002FSKILL.md",[],[],{"skillNo":62,"name":63,"nameZh":6,"description":64,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":65,"categories":66,"keywords":67,"publishedAt":19},"01M7DZSY8XSXBMD53Y5QKRX6CD","wt-steward","Maintains an existing commitments backlog: overdue, about to expire, stalled, and repeatedly-rescheduled commitments, and someday items gone cold. Use when the nightly cron job fires or the user asks to review or tidy their backlog ('review my backlog', 'what's gone stale on my list'). Does not capture or schedule new work — that happens when the user speaks, through wt-capture and wt-schedule.","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fwork-tracker\u002Fskills\u002Fwt-steward\u002FSKILL.md",[],[],{"skillNo":69,"name":70,"nameZh":6,"description":71,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":72,"categories":73,"keywords":74,"publishedAt":19},"01M7DZST40GEJQ67KZSTG1PM93","wt-schedule","Estimates a commitment, links it to its project, and blocks time for it on the user's calendar. Use when the user asks to find time for something — \"when can I do this?\", \"find me two hours for the CFP\", \"block time for that tomorrow\", \"I want to work on it today\" — and immediately after wt-capture files a commitment that needs real hours. Do NOT use for a commitment that needs only a deadline or a reminder, and do NOT use to change what the user committed to — that is wt-update.","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fwork-tracker\u002Fskills\u002Fwt-schedule\u002FSKILL.md",[],[],{"skillNo":76,"name":77,"nameZh":6,"description":78,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":79,"categories":80,"keywords":81,"publishedAt":19},"01M7DZSMTQT1CCMQ3MK984C7FS","wt-capture","Records something the user has said they must do themselves: classifies the commitment as next, waiting, someday or trivial, writes it as an action page in their commitments base, and hands off to wt-schedule when it needs real hours. Use when the user commits to work of their own — \"I need to find time to…\", \"I have to…\", \"remind me to…\", \"help me remember…\", \"put that on my list\", \"I should … at some point\". Do NOT use when the user asks the assistant to do the work (\"write the CFP\", \"research X\", \"email Sam\") — that is a request to act now, not a commitment to track. Do NOT use for thoughts, facts or notes carrying no commitment; those are knowledge and belong to kb-capture.","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fwork-tracker\u002Fskills\u002Fwt-capture\u002FSKILL.md",[],[],{"skillNo":83,"name":84,"nameZh":6,"description":85,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":86,"categories":87,"keywords":88,"publishedAt":19},"01M7DZSG76N1GGR9C82RFK37RJ","work-tracker","Answers questions about the user's own commitments by querying their commitments base, explains the four statuses (next, waiting, someday, done), routes a request to the right work-tracker skill, and states who may change what without asking. Use when the user asks what they should be working on, what is overdue, what they are waiting on, or how their task tracking works, and no narrower work-tracker skill matches. Do NOT use to file a commitment the user just made (that is wt-capture), to find time for one (wt-schedule), or to report progress on one (wt-update).","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fwork-tracker\u002Fskills\u002Fwork-tracker\u002FSKILL.md",[],[],{"skillNo":90,"name":91,"nameZh":6,"description":92,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":93,"categories":94,"keywords":95,"publishedAt":19},"01M7DZSBWW4AQ7J1BTZHCJERTJ","kb","Explains how the user's knowledge bases are laid out and which kb skill handles a given job — the tree, the one queue, expiry, links, base health, maintenance. Use when the user asks how their knowledge base works, what state a base is in, why something was or was not kept, or wants base maintenance run, and no narrower kb skill matches. Do NOT use to file something the user just said (that is kb-capture), to answer a question from stored knowledge (that is kb-recall), or to pick a destination base for a write already in hand (that is kb-route).","https:\u002F\u002Fgithub.com\u002FAlmogBaku\u002Faos\u002Fblob\u002Fmain\u002Ftests\u002Fgolden\u002Fhermes\u002Ffull-install\u002Fhome\u002Frendered\u002Fkb\u002Fskills\u002Fkb\u002FSKILL.md",[],[]]